IEEE Power & Energy Magazine - May/June 2022 - 34

decision-support tool. The ultimate responsibility for the
final decision remains with the operator. ML algorithms
can be extremely fast to estimate, for example, whether
operating points are safe, as ML methods can assess thousands
of scenarios at the same time, whereas conventional
tools can assess only a few. This allows the system operator
to screen an extremely large number of scenarios and determine
the few critical ones in a very short period of time.
Following this step, the operator can then use conventional
techniques to study, in more detail, those few scenarios and
determine what drives them to instability and when.
The approach is similar when we apply ML, and AI in
general, for optimization. Neural networks can determine
a very good estimate for an optimal operating point. We
can then assess, with conventional tools, whether the point
violates any constraints. This way of approaching AI can
drive wider adoption of AI tools for system operation, at
least in the first transition period. But even before this step,
the AI tools need to build trust. If operators and other users
do not trust them, they will not even be used for screening.
The rigorous methods this article describes aim to deliver
the guarantees (through performance certificates) that will
allow users to trust AI tools for safety-critical operations.
Besides the benefits these certificates give us in terms of
trusting AI tools and determining the worst-case performance,
they can help provide better insights into how the
resulting AI model can be improved. The techniques discussed
in this article fall into the broader spectrum of transparent
AI (the results of the AI process are explainable) and
interpretable AI (the accuracy of a model to associate a cause
to an effect), initiatives recently launched in the AI community
to bring a deeper understanding of the inner workings
of ML algorithms.
Several algorithms under this umbrella deliver rigorous
approaches to understand how different features of the problem
at hand affect the performance of the applied neural
network. The algorithms we discuss go a step further and
deliver performance guarantees. To enhance the performance
of AI tools, this article also presents ways to take
advantage of available power system models developed for
decades based on first principles. Through that, neural network
applications can better avoid overlooking critical cases
or giving predictions that are largely at odds with governing
physics. In short, these possibilities are enabled by the following
two approaches:
1) Verification: This approach delivers certificates on
how the neural network will behave, i.e., what its output
will be for all possible inputs. Thus far, researchers
have been assessing the neural network performance
purely statistically, naïvely hoping that testing
an application on a large set of random samples can
accurately capture the behavior of the neural network
model over the whole input space.
2) Physics-informed neural networks: These networks
include the governing physical equations inside a neu34
ieee
power & energy magazine
ral network training procedure. Instead of generating
sets of relevant data in a process to train the model,
which can be computationally intensive, integrate a
physics-based model into the training process and let
the neural network learn from it.
Before diving into these approaches in the next sections, we
first introduce the type of ML tools that are the focus of this
article: neural networks and their training procedure. In the
rest of this article, we use the terms AI and ML interchangeably;
in reality, AI is a more general term that includes all
ML methods and tools.
How Do Neural Networks Work?
Neural networks belong to the most promising group of
AI tools, proving their ability to approximate any function
(universal approximator). They have received considerable
attention in recent years. The two main categories are neural
networks for classification and regression tasks.
Classification neural networks are used in a variety of
applications. In self-driving cars, for example, neural networks
continuously receive pictures of their surroundings
from car cameras and answer simple questions such as, " Is
this a red traffic light? " A classification neural network
shall return a yes or a no, and even if it is not quite sure,
it must choose between the two. For power systems, such a
classification task can be " Will this combination of generator
set points and loads-called an operating point-lead
to a blackout or not? " We can ask this question for many
possible operating points (see the " Possible Inputs " bubble
in Figure 2), and through the neural network, we can obtain
an answer for each point extremely fast. We can label the
resulting regions " safe " (blue dot) or " not safe " (red dots), as
shown in the " Predicted Classifications " bubble in Figure 2.
If we want to assess the predictions of our neural network,
we obviously cannot check for every single operating
point as this requires immense computing resources.
Instead, we sample a fraction of operating points and perform
a conventional security assessment (e.g., run power
flow, time-domain simulations, or an eigenvalue analysis)
and associate a true label with each point, such as truly safe
or truly unsafe. A good neural network will match (i.e., predict)
these true labels as closely as possible.
Instead of a binary response (yes/no), regression neural
networks use almost the same training process as the classification
neural networks but yield a continuous function
value. These networks are widely used, for example, to forecast
load demand or predict the power output of wind turbines
and solar photovoltaics based on weather conditions,
including the movement of clouds, temperature variations,
and others. In a power system operation context, a regression
neural network can predict a cost-optimal operating point
or how the value of frequency evolves after a line outage.
As presented in Figure 2, in the case of frequency evolution,
a neural network predicts a continuous function of the
input to output variables over time (dashed orange line) that
may/june 2022
IEEE Power & Energy Magazine - May/June 2022

Table of Contents for the Digital Edition of IEEE Power & Energy Magazine - May/June 2022
